Most AI memory today is based on things the AI summarizes as “experience.” The problem is that facts can change, and the memory can become wrong.
When a new task starts, the goal, memory, and current facts can conflict. The model may then produce something different from what we expect.
The idea of the anchoring layer is to keep memories tied to the facts they came from, and make it easier to detect when they are no longer valid.
The project is here:
https://github.com/Anchorstate-Lab/GMR
I’d love to hear feedback on the idea.
Zongming•46m ago
Most AI memory today is based on things the AI summarizes as “experience.” The problem is that facts can change, and the memory can become wrong.
When a new task starts, the goal, memory, and current facts can conflict. The model may then produce something different from what we expect.
The idea of the anchoring layer is to keep memories tied to the facts they came from, and make it easier to detect when they are no longer valid.
The project is here:
https://github.com/Anchorstate-Lab/GMR
I’d love to hear feedback on the idea.